7.5M TTS model holds prosody on long sentences: kale-tts-nano f0 std 2.97 st matches 45.7M sibling
NERDDISCO · x · 2026-11-21
- Responding to skepticism that tiny TTS models go flat after a clause or two, the author ran a controlled test: 3 long sentences (commas, a parenthetical, a question), same voice, one take each, default settings.
- nano (7.5M) scores an f0 std of 2.97 semitones, essentially matching their 45.7M tiny (2.93); maya-1 (3B) hits 3.33.
- Charts show per-sentence pitch curves and bars: nano 3.09/2.85/2.97, tiny 3.05/2.89/2.86, maya-1 3.11/3.28/3.61.
- Caveat: this measures pitch movement amount, not whether it moves in the right places — audio demos are the real test.
More from Models
- kale-tts-nano pitch movement holds to the end: 2.83 st early vs 2.87 st late in sentences — NERDDISCO · 2026-11-21
- OVH's Olala Mini 7A1B: 360 tok/s decode on one RTX 5090, 96K context, v2.0 to exceed 100B — JiliJeanlouis · 2026-11-21
- Inkling 2 Mini on a 4090: catches citations near-Haiku level, 2-3x slower than Qwen — strickvl · 2026-11-21
- 6-round eval shows GPT-6.2-sol matching only 11/23 judge outcomes vs 6.1's 21/23 — 0xkarasy · 2026-11-21
- Inkling 2 Mini Now Runs on vLLM, Ending the 'Permanent Underclass' — VoidAsuka · 2026-11-21
- Users report Anthropic classifier firing on any talk of memory with Fable 5 — repligate · 2026-11-21